INTERPRETABLE MICROTEXTURE FEATURE EXTRACTION FOR FABRIC CLASSIFICATION IN TEXTILE RECYCLING SYSTEMS
Authors: MOLCHANOVA, MARYNA; MAZURETS, OLEKSANDR; ZALUTSKA, OLHA; ZAKHARKEVICH, OKSANA; SOBKO, OLENA; BUKHANTSOVA, LIUDMYLA; ORTYNSKYI, VADYM; KOSHEVKO, YULIIA; LOVINSKYI, MAKSYM AND KOZHEVNIKOV, SERHIY
Abstract: In textile waste management and the circular economy, reliable discrimination between natural and synthetic fabrics can support sorting and recycling. Manual inspection is labor-intensive and subjective, while many automated approaches rely on opaque image-based models. This paper presents a reproducible and interpretable method for binary natural-versus-synthetic fabric discrimination from microimages acquired under controlled conditions. The proposed approach represents each image by a fixed numerical profile of 111 interpretable microtexture features rather than raw pixels. Each input image is resized to 256×256 pixels, converted to grayscale, and represented at three automatically defined centered fields of view (96×96, 160×160, and 256×256). The profile includes GLCM descriptors, local binary pattern features, photometric intensity characteristics, Shannon entropy, and gradient-based edge measures. Statistical tests and effect sizes are used for post-hoc characterization and ranking of interpretable indicators; they are not used to select the input variables of the classifier. The source dataset contains 3,107 images (1,547 natural and 1,560 synthetic). The experiment reported in this paper uses a reproducible balanced subset of 400 images, with 200 images per class selected using a fixed random seed (42). The subset is divided by stratified image-level splitting into 280 training images and a held-out test set of 120 images (60 per class). The predictive pipeline is fitted using the training partition only; S63/C15 statistics are reported separately as post-hoc descriptive characterization and do not determine the MLP inputs or the reported test metrics. Of the 111 predefined features, 63 show significant class-distribution differences after FDR correction in this descriptive analysis. A multilayer perceptron trained on the complete 111-feature profile achieves Accuracy = 0.9917 and Macro-F1 = 0.9917 on the held-out 120-image test set. Median imputation and feature standardization are fitted within the training pipeline only. The resulting confusion matrix contains 59 correctly classified natural images, one natural image classified as synthetic, and all 60 synthetic images correctly classified. The threshold-based E14 module is a separate post-hoc interpretation layer; its class ranges and thresholds are calculated from training data and do not constitute or modify the MLP decision boundary. The results demonstrate strong image-level discrimination within the studied acquisition protocol, but they should not be interpreted as specimen-level validation or as a complete industrial textile-identification system. The published image metadata do not preserve physical-specimen identifiers, so specimen-level independence between partitions cannot be verified retrospectively. In addition, photometric features are highly discriminative and may partly encode illumination/reflectance differences. The current task is binary and does not address blends, fibre-content estimation, or multi-class sorting. External validation across independent specimens and imaging conditions is therefore required.
Keywords: Textile recycling; Fabric classification; Microtexture analysis; Interpretable machine learning; Textile waste sorting.
Pages: 2-16
